Summary

• A study by Implement Consulting Group and the ifo Institute estimates €440bn in potential additional annual German value added from generative AI by 2034.
• The model reduces that potential by €155bn under assumptions involving more restrictive access to training data.
• Commissioned by Google and CCIA Europe, the research advocates retaining commercial text and data mining rights alongside transparency, opt-out mechanisms and licensing.

Restrictions on access to training data could reduce Germany’s potential annual economic gains from generative artificial intelligence by €155 billion in 2034, according to a study published by the ifo Institute, adding a substantial but heavily assumption-dependent estimate to Europe’s dispute over copyright and AI development.

The research calculates that generative AI could generate €440 billion in additional annual German value added by 2034 under a relatively favourable set of conditions. A scenario involving more restrictive access to training material reduces that potential by €155 billion, equivalent to slightly more than a third of the theoretical benefit. Neither figure represents a forecast of actual economic output, and the difference should not be interpreted as money Germany has already lost.

Published on 13 October, the study was prepared by Implement Consulting Group in cooperation with the ifo Institute and commissioned by Google and technology industry association CCIA Europe. The sponsors’ interest in the regulation of AI training data is directly relevant to the findings, particularly as the authors recommend preserving broad commercial text and data mining permissions within Europe’s copyright framework.

Oliver Falck, director of ifo’s Center for the Economics of Innovation and Digital Transformation, and Martin Thelle, senior partner at Implement Consulting Group, examine the potential effects of changes to training-data availability on AI capabilities, adoption and research. Their calculations describe possible economic pathways rather than observed causal effects of existing copyright rules.

The assumptions behind the headline figure

Of the €440 billion in potential annual value added identified by the researchers, approximately €330 billion comes from wider AI adoption across the economy. More productive research contributes a further €65 billion in the model, while AI development activity in Germany accounts for the remaining €45 billion.

The study then estimates how more restrictive access to training material might diminish those gains through four channels. Weaker model capabilities account for €75 billion of the proposed difference, slower adoption for €45 billion, reduced research productivity for €25 billion and less European AI development for €10 billion.

These categories describe the model’s allocation of potential effects, not economic losses separately measured in businesses or universities. They also depend on assumptions about how the quality and availability of training data influence the performance of AI systems and how those capabilities subsequently affect economic activity.

The largest component concerns model performance. The researchers argue that restricted access to sufficiently diverse training material could leave European developers with less capable systems, reducing the value that businesses and researchers can obtain from using them.

To illustrate the difference in available data, the study compares what it identifies as Europe’s largest copyright-free dataset, containing approximately two to three trillion tokens, with leading international models trained using between 15 trillion and 36 trillion tokens. These are the authors’ comparisons, while token counts alone do not establish model quality or prove how any individual system would perform with a different training dataset.

The relationship between model capability and economic productivity is also affected by implementation. An organisation may have access to a highly capable AI model without possessing the data infrastructure, skills or internal processes needed to use it productively. Conversely, specialised systems trained on smaller, carefully selected datasets may perform effectively within narrower business applications.

The economic calculation therefore depends on several transitions taking place as assumed: access to data must affect model development, model performance must influence adoption, and that adoption must produce measurable additional output. The study’s numerical precision does not remove uncertainty from those relationships.

Copyright already contains a conditional exception

European law permits certain uses of copyright-protected material for text and data mining under the 2019 Copyright in the Digital Single Market Directive. Article 4 provides an exception covering reproductions and extractions of lawfully accessible works, including commercial uses, provided rights holders have not expressly reserved their rights in an appropriate manner.

The reservation condition distinguishes the exception from an unrestricted entitlement to use copyrighted material. For content publicly available online, the directive envisages machine-readable methods of expressing a reservation, while other arrangements may apply in different circumstances. A separate provision covers qualifying scientific research by research organisations and cultural heritage institutions.

AI developers relying on the commercial exception consequently need to consider whether material is lawfully accessible and whether relevant rights have been reserved. Licensing may be necessary where the exception cannot be relied upon, while different legal questions can arise around the outputs produced by an AI system.

The authors argue that the existing commercial exception should be retained, supported by proportionate transparency obligations, workable international opt-out arrangements and voluntary licensing agreements. They also suggest that some concerns could be addressed through regulation of AI outputs rather than further restrictions on training inputs.

Those recommendations favour a particular balance between the interests of model developers and copyright owners. A less restrictive approach may reduce the expense of assembling training datasets, but the economic consequences for publishers, creative businesses and other rights holders cannot be determined solely by modelling the benefits available to AI companies and their customers.

Licensing provides a means for creators and content owners to receive payment for particular uses of their work. It can also create transaction costs, especially when developers need permission covering extensive collections of material from numerous rights holders. The distribution of those costs and revenues influences the commercial incentives of both technology developers and the industries producing material used in training.

Brussels is examining how opt-outs work

The European Commission has already begun examining the practical operation of copyright reservations for text and data mining. A feasibility study published in July considered an EU-level registry that could help rights holders express opt-outs and make those reservations easier for AI developers to identify.

That research found existing mechanisms fragmented, with machine-readable signals and metadata sometimes implemented inconsistently or lost as content passes between systems. A registry could provide a more durable means of indicating that particular works have been reserved from text and data mining, although the Commission has not committed to creating one.

The proposal concerns the identification of rights reservations rather than replacing licensing markets. Better technical records could potentially help developers establish whether a work is available under the commercial exception while giving rights holders more reliable means of communicating their decisions.

Further policy work is under way. Techopia reported on the Commission’s September consultation examining copyright and generative AI, which runs until 3 November and covers questions about licensing and enforcement alongside other aspects of digital copyright.

That consultation does not itself change European law, and the ifo study should be read as a contribution to the policy debate rather than an assessment of legislation already adopted. Its authors contend that greater regulatory uncertainty can discourage investment, but the announcement does not provide an independently established figure for investment already lost because of the Commission’s review.

Greater clarity could benefit developers and rights holders alike, although their preferred outcomes may differ substantially. Model developers want predictable access to suitable data, while publishers and other creators seek effective control over protected material and commercially viable licensing arrangements.

The study assigns a potential €155 billion annual economic difference to restrictions on training data, but it does not settle how copyright rules should distribute the resulting costs and benefits. Its value to the debate lies in making one set of economic assumptions explicit, while the Commission’s ongoing work will determine whether the existing exception and opt-out mechanisms require changes.